Konstantin (Sber / GigaChat)

#entity

Summary

Engineer on Sber's GigaChat development team, doing R&D on agents. Author of the talk arguing that git-based skills-as-memory are the new memory of AI agents.

Current Understanding

Konstantin brings the model-builder's / R&D lens. His core architectural claim: put tools + data + usage history in one git repo and let the harness decide what to load. He maps the field's evolution (tools → MCP → skills, then agentic-loops) and demonstrates auto-improvement (a weak GigaChat going 1/89 → 11/89 on a benchmark over a weekend).

Evidence

Talk "Git-based skills — the new memory of AI agents": harness definition, two-stage skill loading, git rules + CI back-pressure, personal DNA/trip/HR skills, Hermes curator, Ralph/meta loops — 2026-07-14-skills-based-on-git.
Hackathon record: 7th (Interpress-Ex, infinite-loop harness), 3rd/1st-technical (Snowbase), observed 19/20 top BitGen teams on harnesses.

Concepts: skills-as-memory, harness, evolution-of-agent-tooling, agentic-loops, context-as-scarce-resource
Tools: hermes (skills-first harness he uses/extends), claude-code
Timeline: ai-agent-evolution
Compare: allie-miller (same skills-as-memory idea, business framing), theo-browne (orchestration era)
Comparison: theo-konstantin-allie — three-lens side-by-side (Theo/Konstantin/Allie)

Contradictions / Uncertainty

Pruning windows (30/90 days) and the ">5 tool calls → make a skill" rule are presented as working heuristics from Hermes, not established standards. Status: tentative.
No standards yet for what data to put in a skill or its upper size limit.

Next Questions

How to build world-models / give agents perception (his own "missing piece")?

1.8 KiB Raw Permalink Blame History